Similarity measures for depth estimation

This paper deals with similarity measures for stereoscopic depth estimation. These measures are used for matching of image pairs, which is the first step of the estimation process. We analyze influence of these similarity measures on performance of depth estimation with use of commonly known measures and compare the results with some novel proposals. The performance is judged by increase of quality of view synthesis, which is the main aim of this paper. Experimental results over a variety of moving material demonstrate that considerable gain can be attained without any modifications to estimation core and with tuning of matching stage only. Finally, some guidelines on design of well performing similarity measures are given. For the sake of paper, the whole work is described in context of belief-propagation algorithm, but the results and conclusions apply in general for many other state-of-the art optimization techniques.